Vendoring Policy

  • Vendored libraries MUST not be modified except as required to successfully vendor them.

  • Vendored libraries MUST be released copies of libraries available on PyPI.

  • Vendored libraries MUST be accompanied with LICENSE files.

  • The versions of libraries vendored in pip MUST be reflected in pip/_vendor/vendor.txt.

  • Vendored libraries MUST function without any build steps such as 2to3 or compilation of C code, practically this limits to single source 2.x/3.x and pure Python.

  • Any modifications made to libraries MUST be noted in pip/_vendor/README.rst and their corresponding patches MUST be included tasks/vendoring/patches.

  • Vendored libraries should have corresponding vendored() entries in pip/_vendor/__init__.py.

Rationale

Historically pip has not had any dependencies except for setuptools itself, choosing instead to implement any functionality it needed to prevent needing a dependency. However, starting with pip 1.5, we began to replace code that was implemented inside of pip with reusable libraries from PyPI. This brought the typical benefits of reusing libraries instead of reinventing the wheel like higher quality and more battle tested code, centralization of bug fixes (particularly security sensitive ones), and better/more features for less work.

However, there is several issues with having dependencies in the traditional way (via install_requires) for pip. These issues are:

  • Fragility. When pip depends on another library to function then if for whatever reason that library either isn’t installed or an incompatible version is installed then pip ceases to function. This is of course true for all Python applications, however for every application except for pip the way you fix it is by re-running pip. Obviously, when pip can’t run, you can’t use pip to fix pip, so you’re left having to manually resolve dependencies and installing them by hand.

  • Making other libraries uninstallable. One of pip’s current dependencies is the requests library, for which pip requires a fairly recent version to run. If pip depended on requests in the traditional manner, then we’d either have to maintain compatibility with every requests version that has ever existed (and ever will), OR allow pip to render certain versions of requests uninstallable. (The second issue, although technically true for any Python application, is magnified by pip’s ubiquity; pip is installed by default in Python, in pyvenv, and in virtualenv.)

  • Security. This might seem puzzling at first glance, since vendoring has a tendency to complicate updating dependencies for security updates, and that holds true for pip. However, given the other reasons for avoiding dependencies, the alternative is for pip to reinvent the wheel itself. This is what pip did historically. It forced pip to re-implement its own HTTPS verification routines as a workaround for the Python standard library’s lack of SSL validation, which resulted in similar bugs in the validation routine in requests and urllib3, except that they had to be discovered and fixed independently. Even though we’re vendoring, reusing libraries keeps pip more secure by relying on the great work of our dependencies, and allowing for faster, easier security fixes by simply pulling in newer versions of dependencies.

  • Bootstrapping. Currently most popular methods of installing pip rely on pip’s self-contained nature to install pip itself. These tools work by bundling a copy of pip, adding it to sys.path, and then executing that copy of pip. This is done instead of implementing a “mini installer” (to reduce duplication); pip already knows how to install a Python package, and is far more battle-tested than any “mini installer” could ever possibly be.

Many downstream redistributors have policies against this kind of bundling, and instead opt to patch the software they distribute to debundle it and make it rely on the global versions of the software that they already have packaged (which may have its own patches applied to it). We (the pip team) would prefer it if pip was not debundled in this manner due to the above reasons and instead we would prefer it if pip would be left intact as it is now. The one exception to this, is it is acceptable to remove the pip/_vendor/requests/cacert.pem file provided you ensure that the ssl.get_default_verify_paths().cafile API returns the correct CA bundle for your system. This will ensure that pip will use your system provided CA bundle instead of the copy bundled with pip.

In the longer term, if someone has a portable solution to the above problems, other than the bundling method we currently use, that doesn’t add additional problems that are unreasonable then we would be happy to consider, and possibly switch to said method. This solution must function correctly across all of the situation that we expect pip to be used and not mandate some external mechanism such as OS packages.

Modifications

  • setuptools is completely stripped to only keep pkg_resources

  • pkg_resources has been modified to import its dependencies from pip._vendor

  • packaging has been modified to import its dependencies from pip._vendor

  • html5lib has been modified to import six from pip._vendor and to prefer importing from collections.abc instead of collections.

  • CacheControl has been modified to import its dependencies from pip._vendor

  • requests has been modified to import its other dependencies from pip._vendor and to not load simplejson (all platforms) and pyopenssl (Windows).

Automatic Vendoring

Vendoring is automated via the vendoring.update task (defined in tasks/vendoring/__init__.py) from the content of pip/_vendor/vendor.txt and the different patches in tasks/vendoring/patches/. Launch it via invoke vendoring.update (requires invoke>=0.13.0).

Debundling

As mentioned in the rationale, we, the pip team, would prefer it if pip was not debundled (other than optionally pip/_vendor/requests/cacert.pem) and that pip was left intact. However, if you insist on doing so, we have a semi-supported method (that we don’t test in our CI) and requires a bit of extra work on your end in order to solve the problems described above.

  1. Delete everything in pip/_vendor/ except for pip/_vendor/__init__.py.

  2. Generate wheels for each of pip’s dependencies (and any of their dependencies) using your patched copies of these libraries. These must be placed somewhere on the filesystem that pip can access (pip/_vendor is the default assumption).

  3. Modify pip/_vendor/__init__.py so that the DEBUNDLED variable is True.

  4. Upon installation, the INSTALLER file in pip’s own dist-info directory should be set to something other than pip, so that pip can detect that it wasn’t installed using itself.

  5. (optional) If you’ve placed the wheels in a location other than pip/_vendor/, then modify pip/_vendor/__init__.py so that the WHEEL_DIR variable points to the location you’ve placed them.

  6. (optional) Update the pip_version_check logic to use the appropriate logic for determining the latest available version of pip and prompt the user with the correct upgrade message.

Note that partial debundling is NOT supported. You need to prepare wheels for all dependencies for successful debundling.